7 subsets of image-text pairs feature exhaustive mask annotations for every object instance corresponding to specific Noun Phrases (NPs). These image annotations are verified by a consensus of three human reviewers to provide a high-precision benchmark for promptable concept segmentation (PCS).
Use Cases
- Benchmark promptable concept segmentation models by comparing predicted masks against the human-verified ground truth masks.
- Evaluate zero-shot object detection and segmentation performance using the provided Noun Phrase (NP) text labels as prompts.
- Analyze model performance across different visual domains using the 7 distinct subsets provided in the benchmark.
Strengths
- Contains 7 distinct subsets targeting different annotation domains for diverse visual contexts.
- Features exhaustive mask annotations for every object instance matching the provided Noun Phrase (NP) labels.
- Every annotation is multi-reviewed and agreed upon by 3 human annotators to ensure label precision.